Design and Optimization of HXDSP Heterogeneous Computing Framework
NING Cheng-ming1,CAI Heng-yu1,ZHENG Qi-long1,2,GENG Rui3
1(School of Computer Science,University of Science and Technology of China,Hefei 230026,China)2(Anhui Province Key Laboratory of High Performance Computing,University of Science and Technology of China,Hefei 230026,China)3(Anhui Siliepoch Technology Co.,Ltd,Hefei 230088,China)
Abstract：Due to the performance bottleneck that CPU faces when processing massive amounts of data,heterogeneous parallel computing based on heterogeneous systems has become one of the research hotspots in the field of parallel computing.HXDSP is a DSP chip independently developed by China Electronics Technology Group 38,which can meet the needs of a variety of high-performance computing fields.In order to give full play to the computing power of HXDSP and the parallel computing power of multiple HXDSP devices,this paper designs a heterogeneous computing framework for HXDSP based on OpenCL.In this paper,different OpenCL device mapping methods are designed according to the characteristics of computing tasks,so that the HXDSP heterogeneous computing framework can adapt to different application scenarios,make full use of HXDSP computing resources and save energy consumption.Based on the design of the HXDSP heterogeneous computing framework,this article optimizes the HXDSP heterogeneous computing applications,including data transmission optimization and data access optimization.Finally,this paper analyzes the program performance of the HXDSP heterogeneous computing framework through experiments,and verifies the feasibility and effectiveness of the HXDSP heterogeneous computing framework.
2022, Vol. 43 
